EGU23 7343 Presentation

Streamflow-based snow mass reconstruction
using spatial melt patterns from synthetic snow cover maps

Pau Wiersma (pau.wiersma@unil.ch), Fatemeh Zakeri & Grégoire Mariéthoz
Graphical abstract Introduction Models & Data
Methods Results Conclusion & Outlook

Temperature + Setting Accumulation Melt factor

Precipitation Daily 30m snow cover maps
Snowfall correction
10.000x + S1
Hydrological Model No snowfall correction Constant, uniform

(1km resolution)
2D melt factor adjustment

S2
Snowfall correction Constant, uniform
10.000 prior daily snow mass maps

S3
Snowfall correction Constant, seasonal

Streamflow observations
S4
Evaluation against
100 posterior daily snow mass maps Snowfall correction
snow observations Distributed, seasonal
Model & Data:

• Wflow_sbm hydrological model
• Resampled Landsat daily synthetic snow cover maps
• Thur-Jonschwil catchment in Eastern Switzerland
(493 km2) with 2 discharge and 4 snow depth stations
Methodology:
• Per setting 500 calibration runs using only streamflow
observations at Jonschwil
• 10 best runs are kept for evaluation on
• Mogelsberg streamflow
• In-situ gap-filled snow depth to snow mass
observations
• MODIS snow cover (500m)
• Setting 4:
Distributed melt factor =
uniform melt factor * 1km gridded snow cover fraction
Results:
• Snowfall correction improves all results
• Replacing uniform with distributed melt factor leads to:
• Slightly improved streamflow
• Similar snow mass performance
• Reduced spatial snow cover performance, except in
early winter
Motivation Temperature +
Precipitation
Snowfall correction
• By tweaking the snow accumulation and melt configurations of a 10.000x +
distributed hydrological model, we can simulate an infinite number of
spatio-temporal snow water equivalent (SWE) distributions Hydrological Model
(1km resolution)
• All these SWE distributions lead to a different streamflow response
after having been melted and routed by the hydrological model 2D melt factor adjustment
• Comparison against streamflow observations allows us determine
which SWE distributions were realistic and which ones should be
discarded
• This finally gives us a posterior ensemble of catchment-wide SWE 10.000 prior daily snow mass maps
evolutions over each winter
• However, these hydrological models often use simplifications that
prevent a realistic representation of spatial melt patterns, which leads
Streamflow observations
to difficulties in the SWE-streamflow translation
100 posterior daily snow mass maps

Problem statement
Snow water equivalent Snow melt response
• In reality, the lower the snow cover fractions, the
less area contributes to the snow melt and therefore
the smoother the melt response will be Reality
• In hydrological models, a grid cell can often only be
either fully snow covered or fully snow depleted
• Since the entire grid cell area is contributing to the
snow melt until all the snow has disappeared, the
simulated melt is likely too abrupt Hydrological Model
• To prevent this behaviour, remotely sensed snow
cover can inform the model on the subgrid snow
cover fraction and help to smooth out the simulated
Hydrological model
melt +
• But what about days without remotely sensed snow Remotely sensed
cover observations? snow cover fraction
+
Proposal
• Recently, multiple studies have attempted to create a daily time

series of high-resolution snow cover maps by downscaling
MODIS or resampling Landsat/Sentinel-2 imagery Real snow cover maps
• Zakeri & Mariéthoz (in preparation), have developed a
methodology that can run independently from MODIS imagery
and has the potential to synthetize snow cover maps before 2000
and into the future
• In this study, we assess the added benefit of using these synthetic
snow cover maps to inform the hydrological model on the per 1950 2100
pixel snow cover fraction and consequently the per pixel melt
factor reduction
• We do this by running the streamflow-based SWE reconstruction Synthetic snow cover maps
with and without the synthetic snow cover maps and consequently
evaluating the results using in-situ observations and MODIS snow
cover maps
• The study is carried out between 1999-2008 in the Thur-Jonschwil
catchment in eastern Switzerland using the wflow_sbm
hydrological model
Hydrological Model
• Wflow_sbm (van Verseveld et al., in review)

• Snow module:
• Gridded version of HBV-96
• Constant and uniform degree day melt factor
• Liquid water content + refreezing
• Partial rain/snowfall around 0° C
• Lateral gravitational transport
• Julia source code run within the python wrapper
pywflow (Schilperoort, B. (2023). github.com/eWaterCycle/pywflow)
• 1km resolution
• We use the station-based gridded daily TabsD
(temperature) and RhiresD (precipitation) products of
MeteoSwiss as forcing (2km resolution)
Test catchment
• Thur-Jonschwil in Eastern Switzerland (493 km 2)

• Intermittent snow cover at lower elevations
• Mean peak SWE of 650mm at Säntis (2502m)
• Discharge measurements (BAFU)
• Jonschwil (1966-2023)
• Mogelsberg (88 km2, 1972 – 2023)
• Snow depth measurements
• Schwägalp (SLF, 1348m)
• Wildhaus (MeteoSwiss, 998m)
• Mogelsberg (MeteoSwiss, 780m)
• St Peterzell (MeteoSwiss, 700m)
• The snow depth data have been gap-filled and
converted to SWE by SLF
Methods (1/2) Setting Accumulation Ablation

• 4 model settings of increasing complexity
• In S4, the snow cover fraction is calculated for each S1
1km grid cell using the fraction of 30m synthetic binary
snow cover maps No snowfall correction Constant, uniform
• This snow cover fraction is consequently
multiplied by the basin-wide melt factor to obtain
S2
the distributed melt factor
• Whenever there is still SWE in the hydrological Snowfall correction Constant, uniform
model but the 1km snow cover fraction is zero it is
converted to 100% to accelerate the meltout and
prevent SWE persistance S3
S4
Snowfall correction Distributed, seasonal

Methods (2/2) Setting Accumulation Ablation

• For each model setting:
• 500 calibration runs using the ROPE optimization S1
algorithm (Bardossy et al., 2008) within SPOTPY (Houska
No snowfall correction Constant, uniform
et al., 2015) over the period 1999-2008
• Keep the 10 runs with the highest KGE for
evaluation S2
• For calibration we use only the streamflow at Jonschwil
For evaluation we use: Snowfall correction Constant, uniform
• Streamflow at Mogelsberg
• Modeled SWE at snow stations
• MODIS snow cover upscaled to 1km S3
• For all setting two sensitive soil parameters are included
in the calibration
• For S1+2 we calibrate the static melt factor, while for
S3+4 we calibrate both the minimum and maximum
between which the melt factor fluctuates S4
Snowfall correction Distributed, seasonal

Results - streamflow evaluation
• Applying the snowfall increases KGE performance

• No significant improvement from making the melt factor
vary in time or space
Results - Snow station evaluation
• Reduced underestimation with the distributed melt factor

(S4) at the highest station (Schwägalp)
• Frequent overestimation and reduced performance at other
stations
Results - Snow station evaluation
• Reduced underestimation with the distributed melt factor

(S4) at the highest station (Schwägalp)
• Frequent overestimation and reduced performance at other
stations
Results - MODIS snow cover evaluation
• Despite the added information by the synthetic snow cover

maps, the overall performance of S4 is reduced compared to S3
• Particularly in late spring S4 contains many false positives (i.e.
simulated meltout date is too late)
• The opposite is the case in early winter, when S4 performs better
due to less false negatives (i.e. early winter meltouts are
correctly prevented)
Results - MODIS snow cover evaluation
• Despite the added information by the synthetic snow cover

maps, the overall performance of S4 is reduced compared to S3
• Particularly in late spring S4 contains many false positives (i.e.
simulated meltout date is too late)
• The opposite is the case in early winter, when S4 performs better
due to less false negatives (i.e. early winter meltouts are
correctly prevented)
Preliminary conclusions Next steps
• We performed streamflow-based snow mass • Assess the sensivity to the synthetic snow cover maps
reconstruction using 4 configurations of increasing • Repeat setting 4 with MODIS snow cover fraction
complexity • Include more snow-dominated catchments
• Applying a snow cover correction (S1-2) improved • Compare multi-winter vs single winter reconstruction
streamflow, SWE and spatial snow cover results in • Explore other ways of translating the snow cover
nearly all cases fraction information to a distributed melt factor in the
• Going from a temporally constant to a seasonal melt hydrological model to include:
factor (S2-3) only lead to a significant improvement in • Vegetation
spatial snow cover results • Snowpack age
• Including synthetic spatial snow cover fraction maps to • Shadow
go from a spatially uniform to a spatially distributed melt
factor (S3-4) surprisingly lead to reduced spatial snow
cover performance, except in early winter. Streamflow
and SWE modeling performance did not change
significantly
References
• Bárdossy, A. and S. K. Singh. 2008. Robust estimation of hydrological model parameters. Hydrology and Earth
System Sciences 12, Nr. 6: 1273–1283. doi:10.5194/hess-12-1273-2008, .
• Houska, Tobias, Philipp Kraft, Alejandro Chamorro-Chavez and Lutz Breuer. 2015. SPOTting Model Parameters
Using a Ready-Made Python Package. PLoS ONE 10, Nr. 12: e0145180. doi:10.1371/journal.pone.0145180, .
• Schilperoort, B. (2023). pywflow (Version 0.1.0) github.com/eWaterCycle/pywflow
• Verseveld, Willem J. van, Albrecht H. Weerts, Martijn Visser, Joost Buitink, Ruben O. Imhoff, Hélène Boisgontier,
Laurène Bouaziz, et al. 2022. Wflow_sbm v0.6.1, a spatially distributed hydrologic model: from global data to local
applications. Geoscientific Model Development Discussions 2022: 1–52. doi:10.5194/gmd-2022-182, .
Thanks to Tobias Jonas from SLF for providing the SWE data at the four snow stations.
ROPE optimization algorithm (Bardossy et al., 2008)

• Latin hypercube for the first half of runs
• Keep 20% best results and start latin hypercube for next 1/6
• Repeat twice for a total of 3 resampling subsets

EGU23 7343 Presentation

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

EGU23 7343 Presentation

Uploaded by

Copyright:

Available Formats

Streamflow-based snow mass reconstruction

using spatial melt patterns from synthetic snow cover maps

Graphical abstract Introduction Models & Data

Methods Results Conclusion & Outlook

Temperature + Setting Accumulation Melt factor

Hydrological Model No snowfall correction Constant, uniform

2D melt factor adjustment

Snowfall correction Constant, uniform

10.000 prior daily snow mass maps

Snowfall correction Constant, seasonal

Model & Data:

100 posterior daily snow mass maps

• Recently, multiple studies have attempted to create a daily time

• Wflow_sbm (van Verseveld et al., in review)

• Thur-Jonschwil in Eastern Switzerland (493 km 2)

Methods (1/2) Setting Accumulation Ablation

Snowfall correction Distributed, seasonal

Methods (2/2) Setting Accumulation Ablation

Snowfall correction Distributed, seasonal

Results - streamflow evaluation

• Applying the snowfall increases KGE performance

Results - Snow station evaluation

• Reduced underestimation with the distributed melt factor

Results - Snow station evaluation

• Reduced underestimation with the distributed melt factor

Results - MODIS snow cover evaluation

• Despite the added information by the synthetic snow cover

Results - MODIS snow cover evaluation

• Despite the added information by the synthetic snow cover

Preliminary conclusions Next steps

ROPE optimization algorithm (Bardossy et al., 2008)

You might also like